[Bugfix] Fix model run _npu_flash_attention hang issue (#4410)

Fix model run _npu_flash_attention in _forward_prefill_no_cache hang
issue, it was caused by wrong attention mask dtype.
### How was this patch tested?
Yes, tesed on Qwen2.5-VL and Qwen2.5-Omni

- vLLM version: v0.11.0
- vLLM main:
2918c1b49c

Signed-off-by: Ting FU <futing10@huawei.com>

tests/ut/attention/test_attention_mask.py

@@ -74,10 +74,11 @@ class TestAttentionMaskBuilder(TestBase):
         attn_mask = attention_mask_builder.get_attn_mask(
             max_seq_len=2048, dtype=torch.float16, device=torch.device("cpu"))
         self.assertEqual(attn_mask.shape, (2048, 2048))
-        self.assertEqual(attn_mask[0][-1], torch.tensor(True))
+        self.assertEqual(attn_mask[0][-1],
-        self.assertEqual(attention_mask_builder._seq_len_cached, 1024)
+                         torch.tensor(float("-inf"), dtype=torch.float16))
+        self.assertEqual(attention_mask_builder._seq_len_cached, 2048)
         self.assertEqual(attention_mask_builder.attn_mask_cache.shape,
-                         (1024, 1024))
+                         (2048, 2048))
         self.assertEqual(attention_mask_builder.attn_mask_cache[0][-1],
                          torch.tensor(float("-inf"), dtype=torch.float16))
 

vllm_ascend/attention/attention_mask.py

@@ -67,8 +67,6 @@ class AttentionMaskBuilder:
 
     def get_attn_mask(self, max_seq_len: int, dtype: torch.dtype,
                       device: torch.device):
-        if max_seq_len == 2048:
-            return self.chunked_prefill_attn_mask.to(torch.bool)
         self._update_attn_cache(max_seq_len, dtype)
         return self.attn_mask_cache[:max_seq_len, :max_seq_len].contiguous(
         ).to(device, non_blocking=True)

vllm_ascend/worker/model_runner_v1.py

@@ -991,8 +991,8 @@ class NPUModelRunner(LoRAModelRunnerMixin):
                 max_seq_len, self.dtype, self.device)
         # Prefill with cache hit.
         elif attn_state == AscendAttentionState.PrefillCacheHit:
-            return self.attn_mask_builder.get_attn_mask(
+            return self.attn_mask_builder.get_splitfuse_attn_mask().to(
-                2048, self.dtype, self.device)
+                torch.bool)
         # Decode-only situation.
         else:
             return None